1use std::{
2 error::Error,
3 fmt::{self, Debug},
4 path::Path,
5 sync::{Arc, Mutex},
6 time::Duration,
7};
8
9use crate::{
10 ToolMetrics,
11 assertions::{AssertionsReport, RanAssertion, RanAssertionResult},
12};
13use agent::{ContextLoadResult, Thread, ThreadEvent};
14use anyhow::{Result, anyhow};
15use async_trait::async_trait;
16use buffer_diff::DiffHunkStatus;
17use collections::HashMap;
18use futures::{FutureExt as _, StreamExt, channel::mpsc, select_biased};
19use gpui::{App, AppContext, AsyncApp, Entity};
20use language_model::{LanguageModel, Role, StopReason};
21
22pub const THREAD_EVENT_TIMEOUT: Duration = Duration::from_secs(60 * 2);
23
24#[async_trait(?Send)]
25pub trait Example {
26 fn meta(&self) -> ExampleMetadata;
27 async fn conversation(&self, cx: &mut ExampleContext) -> Result<()>;
28 fn diff_assertions(&self) -> Vec<JudgeAssertion> {
29 Vec::new()
30 }
31 fn thread_assertions(&self) -> Vec<JudgeAssertion> {
32 Vec::new()
33 }
34}
35
36#[derive(Clone, Debug)]
37pub struct JudgeAssertion {
38 pub id: String,
39 pub description: String,
40}
41
42#[derive(Clone, Debug)]
43pub struct ExampleMetadata {
44 pub name: String,
45 pub url: String,
46 pub revision: String,
47 pub language_server: Option<LanguageServer>,
48 pub max_assertions: Option<usize>,
49}
50
51#[derive(Clone, Debug)]
52pub struct LanguageServer {
53 pub file_extension: String,
54 pub allow_preexisting_diagnostics: bool,
55}
56
57impl ExampleMetadata {
58 pub fn repo_name(&self) -> String {
59 self.url
60 .split('/')
61 .next_back()
62 .unwrap_or(&"")
63 .trim_end_matches(".git")
64 .into()
65 }
66}
67
68pub struct FailedAssertion(pub String);
69
70impl fmt::Debug for FailedAssertion {
71 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
72 write!(f, "Assertion failure: {}", self.0)
73 }
74}
75
76impl fmt::Display for FailedAssertion {
77 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
78 write!(f, "{}", self.0)
79 }
80}
81
82impl Error for FailedAssertion {}
83
84pub struct ExampleContext {
85 meta: ExampleMetadata,
86 log_prefix: String,
87 agent_thread: Entity<agent::Thread>,
88 app: AsyncApp,
89 model: Arc<dyn LanguageModel>,
90 pub assertions: AssertionsReport,
91 pub tool_metrics: Arc<Mutex<ToolMetrics>>,
92}
93
94impl ExampleContext {
95 pub fn new(
96 meta: ExampleMetadata,
97 log_prefix: String,
98 agent_thread: Entity<agent::Thread>,
99 model: Arc<dyn LanguageModel>,
100 app: AsyncApp,
101 ) -> Self {
102 let assertions = AssertionsReport::new(meta.max_assertions);
103
104 Self {
105 meta,
106 log_prefix,
107 agent_thread,
108 assertions,
109 model,
110 app,
111 tool_metrics: Arc::new(Mutex::new(ToolMetrics::default())),
112 }
113 }
114
115 pub fn push_user_message(&mut self, text: impl ToString) {
116 self.app
117 .update_entity(&self.agent_thread, |thread, cx| {
118 thread.insert_user_message(
119 text.to_string(),
120 ContextLoadResult::default(),
121 None,
122 cx,
123 );
124 })
125 .unwrap();
126 }
127
128 pub fn assert(&mut self, expected: bool, message: impl ToString) -> Result<()> {
129 let message = message.to_string();
130 self.log_assertion(
131 if expected {
132 Ok(())
133 } else {
134 Err(anyhow::Error::from(FailedAssertion(message.clone())))
135 },
136 message,
137 )
138 }
139
140 pub fn assert_some<T>(&mut self, option: Option<T>, message: impl ToString) -> Result<T> {
141 let message = message.to_string();
142 self.log_assertion(
143 match option {
144 Some(value) => Ok(value),
145 None => Err(anyhow::Error::from(FailedAssertion(message.clone()))),
146 },
147 message,
148 )
149 }
150
151 #[allow(dead_code)]
152 pub fn assert_eq<T: PartialEq + Debug>(
153 &mut self,
154 left: T,
155 right: T,
156 message: impl ToString,
157 ) -> Result<()> {
158 let message = message.to_string();
159 self.log_assertion(
160 if left == right {
161 Ok(())
162 } else {
163 println!(
164 "{}{}",
165 self.log_prefix,
166 pretty_assertions::Comparison::new(&left, &right)
167 );
168 Err(anyhow::Error::from(FailedAssertion(message.clone())))
169 },
170 message,
171 )
172 }
173
174 fn log_assertion<T>(&mut self, result: Result<T>, message: String) -> Result<T> {
175 if let Some(max) = self.meta.max_assertions {
176 if self.assertions.run_count() > max {
177 return Err(anyhow!(
178 "More assertions were run than the stated max_assertions of {}",
179 max
180 ));
181 }
182 }
183
184 self.assertions.ran.push(RanAssertion {
185 id: message.clone(),
186 result: Ok(RanAssertionResult {
187 analysis: None,
188 passed: result.is_ok(),
189 }),
190 });
191
192 if result.is_ok() {
193 println!("{}✅ {}", self.log_prefix, message);
194 } else {
195 println!("{}❌ {}", self.log_prefix, message);
196 }
197
198 result
199 }
200
201 pub async fn run_to_end(&mut self) -> Result<Response> {
202 self.run_turns(u32::MAX).await
203 }
204
205 pub async fn run_turn(&mut self) -> Result<Response> {
206 self.run_turns(1).await
207 }
208
209 pub async fn run_turns(&mut self, iterations: u32) -> Result<Response> {
210 let (mut tx, mut rx) = mpsc::channel(1);
211
212 let tool_metrics = self.tool_metrics.clone();
213 let log_prefix = self.log_prefix.clone();
214 let _subscription = self.app.subscribe(
215 &self.agent_thread,
216 move |thread, event: &ThreadEvent, cx| match event {
217 ThreadEvent::ShowError(thread_error) => {
218 tx.try_send(Err(anyhow!(thread_error.clone()))).ok();
219 }
220 ThreadEvent::Stopped(reason) => match reason {
221 Ok(StopReason::EndTurn) => {
222 tx.close_channel();
223 }
224 Ok(StopReason::ToolUse) => {
225 if thread.read(cx).remaining_turns() == 0 {
226 tx.close_channel();
227 }
228 }
229 Ok(StopReason::MaxTokens) => {
230 tx.try_send(Err(anyhow!("Exceeded maximum tokens"))).ok();
231 }
232 Err(err) => {
233 tx.try_send(Err(anyhow!(err.clone()))).ok();
234 }
235 },
236 ThreadEvent::StreamedAssistantText(_, _)
237 | ThreadEvent::StreamedAssistantThinking(_, _)
238 | ThreadEvent::UsePendingTools { .. } => {}
239 ThreadEvent::ToolFinished {
240 tool_use_id,
241 pending_tool_use,
242 ..
243 } => {
244 thread.update(cx, |thread, _cx| {
245 if let Some(tool_use) = pending_tool_use {
246 let mut tool_metrics = tool_metrics.lock().unwrap();
247 if let Some(tool_result) = thread.tool_result(&tool_use_id) {
248 let message = if tool_result.is_error {
249 format!("✖︎ {}", tool_use.name)
250 } else {
251 format!("✔︎ {}", tool_use.name)
252 };
253 println!("{log_prefix}{message}");
254 tool_metrics
255 .insert(tool_result.tool_name.clone(), !tool_result.is_error);
256 } else {
257 let message =
258 format!("TOOL FINISHED WITHOUT RESULT: {}", tool_use.name);
259 println!("{log_prefix}{message}");
260 tool_metrics.insert(tool_use.name.clone(), true);
261 }
262 }
263 });
264 }
265 ThreadEvent::InvalidToolInput { .. } => {
266 println!("{log_prefix} invalid tool input");
267 }
268 ThreadEvent::ToolConfirmationNeeded => {
269 panic!(
270 "{}Bug: Tool confirmation should not be required in eval",
271 log_prefix
272 );
273 }
274 ThreadEvent::StreamedCompletion
275 | ThreadEvent::MessageAdded(_)
276 | ThreadEvent::MessageEdited(_)
277 | ThreadEvent::MessageDeleted(_)
278 | ThreadEvent::SummaryChanged
279 | ThreadEvent::SummaryGenerated
280 | ThreadEvent::ReceivedTextChunk
281 | ThreadEvent::StreamedToolUse { .. }
282 | ThreadEvent::CheckpointChanged
283 | ThreadEvent::UsageUpdated(_)
284 | ThreadEvent::CancelEditing => {
285 tx.try_send(Ok(())).ok();
286 if std::env::var("ZED_EVAL_DEBUG").is_ok() {
287 println!("{}Event: {:#?}", log_prefix, event);
288 }
289 }
290 },
291 );
292
293 let model = self.model.clone();
294
295 let message_count_before = self.app.update_entity(&self.agent_thread, |thread, cx| {
296 thread.set_remaining_turns(iterations);
297 thread.send_to_model(model, None, cx);
298 thread.messages().len()
299 })?;
300
301 loop {
302 select_biased! {
303 result = rx.next() => {
304 if let Some(result) = result {
305 result?;
306 } else {
307 break;
308 }
309 }
310 _ = self.app.background_executor().timer(THREAD_EVENT_TIMEOUT).fuse() => {
311 return Err(anyhow!("Agentic loop stalled - waited {:?} without any events", THREAD_EVENT_TIMEOUT));
312 }
313 }
314 }
315
316 let messages = self.app.read_entity(&self.agent_thread, |thread, cx| {
317 let mut messages = Vec::new();
318 for message in thread.messages().skip(message_count_before) {
319 messages.push(Message {
320 _role: message.role,
321 text: message.to_string(),
322 tool_use: thread
323 .tool_uses_for_message(message.id, cx)
324 .into_iter()
325 .map(|tool_use| ToolUse {
326 name: tool_use.name.to_string(),
327 value: tool_use.input,
328 })
329 .collect(),
330 });
331 }
332 messages
333 })?;
334
335 let response = Response::new(messages);
336
337 Ok(response)
338 }
339
340 pub fn edits(&self) -> HashMap<Arc<Path>, FileEdits> {
341 self.agent_thread
342 .read_with(&self.app, |thread, cx| {
343 let action_log = thread.action_log().read(cx);
344 HashMap::from_iter(action_log.changed_buffers(cx).into_iter().map(
345 |(buffer, diff)| {
346 let snapshot = buffer.read(cx).snapshot();
347
348 let file = snapshot.file().unwrap();
349 let diff = diff.read(cx);
350 let base_text = diff.base_text().text();
351
352 let hunks = diff
353 .hunks(&snapshot, cx)
354 .map(|hunk| FileEditHunk {
355 base_text: base_text[hunk.diff_base_byte_range.clone()].to_string(),
356 text: snapshot
357 .text_for_range(hunk.range.clone())
358 .collect::<String>(),
359 status: hunk.status(),
360 })
361 .collect();
362
363 (file.path().clone(), FileEdits { hunks })
364 },
365 ))
366 })
367 .unwrap()
368 }
369
370 pub fn agent_thread(&self) -> Entity<Thread> {
371 self.agent_thread.clone()
372 }
373}
374
375impl AppContext for ExampleContext {
376 type Result<T> = anyhow::Result<T>;
377
378 fn new<T: 'static>(
379 &mut self,
380 build_entity: impl FnOnce(&mut gpui::Context<T>) -> T,
381 ) -> Self::Result<Entity<T>> {
382 self.app.new(build_entity)
383 }
384
385 fn reserve_entity<T: 'static>(&mut self) -> Self::Result<gpui::Reservation<T>> {
386 self.app.reserve_entity()
387 }
388
389 fn insert_entity<T: 'static>(
390 &mut self,
391 reservation: gpui::Reservation<T>,
392 build_entity: impl FnOnce(&mut gpui::Context<T>) -> T,
393 ) -> Self::Result<Entity<T>> {
394 self.app.insert_entity(reservation, build_entity)
395 }
396
397 fn update_entity<T, R>(
398 &mut self,
399 handle: &Entity<T>,
400 update: impl FnOnce(&mut T, &mut gpui::Context<T>) -> R,
401 ) -> Self::Result<R>
402 where
403 T: 'static,
404 {
405 self.app.update_entity(handle, update)
406 }
407
408 fn read_entity<T, R>(
409 &self,
410 handle: &Entity<T>,
411 read: impl FnOnce(&T, &App) -> R,
412 ) -> Self::Result<R>
413 where
414 T: 'static,
415 {
416 self.app.read_entity(handle, read)
417 }
418
419 fn update_window<T, F>(&mut self, window: gpui::AnyWindowHandle, f: F) -> Result<T>
420 where
421 F: FnOnce(gpui::AnyView, &mut gpui::Window, &mut App) -> T,
422 {
423 self.app.update_window(window, f)
424 }
425
426 fn read_window<T, R>(
427 &self,
428 window: &gpui::WindowHandle<T>,
429 read: impl FnOnce(Entity<T>, &App) -> R,
430 ) -> Result<R>
431 where
432 T: 'static,
433 {
434 self.app.read_window(window, read)
435 }
436
437 fn background_spawn<R>(
438 &self,
439 future: impl std::future::Future<Output = R> + Send + 'static,
440 ) -> gpui::Task<R>
441 where
442 R: Send + 'static,
443 {
444 self.app.background_spawn(future)
445 }
446
447 fn read_global<G, R>(&self, callback: impl FnOnce(&G, &App) -> R) -> Self::Result<R>
448 where
449 G: gpui::Global,
450 {
451 self.app.read_global(callback)
452 }
453}
454
455#[derive(Debug)]
456pub struct Response {
457 messages: Vec<Message>,
458}
459
460impl Response {
461 pub fn new(messages: Vec<Message>) -> Self {
462 Self { messages }
463 }
464
465 pub fn expect_tool(
466 &self,
467 tool_name: &'static str,
468 cx: &mut ExampleContext,
469 ) -> Result<&ToolUse> {
470 let result = self.messages.iter().find_map(|msg| {
471 msg.tool_use
472 .iter()
473 .find(|tool_use| tool_use.name == tool_name)
474 });
475 cx.assert_some(result, format!("called `{}`", tool_name))
476 }
477
478 #[allow(dead_code)]
479 pub fn tool_uses(&self) -> impl Iterator<Item = &ToolUse> {
480 self.messages.iter().flat_map(|msg| &msg.tool_use)
481 }
482
483 pub fn texts(&self) -> impl Iterator<Item = String> {
484 self.messages.iter().map(|message| message.text.clone())
485 }
486}
487
488#[derive(Debug)]
489pub struct Message {
490 _role: Role,
491 text: String,
492 tool_use: Vec<ToolUse>,
493}
494
495#[derive(Debug)]
496pub struct ToolUse {
497 pub name: String,
498 value: serde_json::Value,
499}
500
501impl ToolUse {
502 pub fn parse_input<Input>(&self) -> Result<Input>
503 where
504 Input: for<'de> serde::Deserialize<'de>,
505 {
506 serde_json::from_value::<Input>(self.value.clone()).map_err(|err| anyhow!(err))
507 }
508}
509
510#[derive(Debug, Eq, PartialEq)]
511pub struct FileEdits {
512 pub hunks: Vec<FileEditHunk>,
513}
514
515#[derive(Debug, Eq, PartialEq)]
516pub struct FileEditHunk {
517 pub base_text: String,
518 pub text: String,
519 pub status: DiffHunkStatus,
520}
521
522impl FileEdits {
523 pub fn has_added_line(&self, line: &str) -> bool {
524 self.hunks.iter().any(|hunk| {
525 hunk.status == DiffHunkStatus::added_none()
526 && hunk.base_text.is_empty()
527 && hunk.text.contains(line)
528 })
529 }
530}